The traditional approach to speech recognition with several intermediate parts is exemplified as follows:

$Audio:\left(input\right)->feature:extraction->phoneme:detection->word:composition->text:transcript:\left(output\right)$

The end-to-end approach replaces the intermediate chain into a single deep neural network, allowing the use of a single optimization criterion for enhancing the system:

$Audio:\left(input\right)-\left(Neural:Network\right)->text:transcript:\left(output\right)$